Driver Fatigue Detection Based On Deep Learning

Driver fatigue driving is an important cause of traffic accidents.The detection and early warning of the driver’s driving state can effectively avoid the occurrence of traffic accidents.Driver fatigue detection has become a research hotspot in the current traffic field,but the accuracy of the current detection method remains at a low level.This thesis researches driver fatigue detection algorithm based on deep learning,combines face and feature points to identify driver status through convolutional neural network.The main work of the thesis is as follows:(1)Propose a face and feature point detection algorithm based on MTCNN(Multi-Task Convolutional Neural Networks,MTCNN)to achieve accurate detection of driver’s face area and relevant feature points positioning under real driving environment.The algorithm adds training feature points,improves the accuracy of network feature extraction and the feasibility of head pose judgment;replaces the network’s excitation function and adds a BN layer to increase the network’s computing speed;meanwhile,in the O-Net network structure Add a multi-scale pooling layer to further improve the network model’s ability to describe features at different scales.(2)Propose an eye and mouth state recognition model based on deep separable convolutional neural network.The features of eyes and mouth are extracted and input to the deep separable convolutional neural network model to judge the feature state,which greatly improves the detection efficiency.At the same time,the head posture is judged according to the face feature points,which is used for the fatigue judgment of the subsequent multi-feature fusion.(3)Propose a multi-feature fusion driver fatigue judgment algorithm.According to the recognized eyes,mouth state and head posture,three characteristic indexes of PERCLOS value,mouth opening rate MCR and head non-face rate HDR are set,and multi-features are combined to comprehensively judge the driver’s fatigue state.(4)The face and feature point detection algorithm,the eye and mouth state recognition algorithm and the multi-feature fusion fatigue judgment algorithm were tested and tested respectively.For face and feature point detection algorithms,the MTCNN algorithm before and after improvement is compared on the FDDB data set.The results show that the algorithm in this thesis is superior to the original algorithm in detection speed and accuracy.Aiming at the eyes and mouth state recognition algorithm,the deep separable convolutional network model in this thesis is compared with the conventional convolutional model and the traditional detection algorithm.Experiments show that the time-consuming and accurate rate of this model is significantly better than the comparison algorithm.Aiming at the driver’s fatigue judgment algorithm,this thesis compares it with several other detection algorithms on the YawDD dataset and the self-made simulation driving process video of the volunteers of the research group.Experiments show that the algorithm in this thesis has improved the accuracy and recall rate.reaching 96.22%and 98.08%respectively.